Online fraud prevention using genetic algorithm solution
Online fraud prevention including receiving a rules set to detect fraud, mapping the rules set to a data set, mapping success data to members of the rules set, filtering the members of the rules set, and ordering members of the data set by giving priority to those members of the data set with a greater probability for being fraudulent based upon the success data of each member of the rule set in detecting fraud. Further, a receiver coupled to an application server to receive a rules set to detect fraud, and a server coupled to the application server, to map the rules set to a data set, and to map the success data to each members of the rules set. The server is used to order the various members of the data set by giving priority to those members of the data set with a greatest probability for being fraudulent.
Latest eBay Patents:
This application is a continuation of U.S. application Ser. No. 12/638,942, filed Dec. 15, 2009, now U.S. Pat. No. 7,840,520 entitled “Online Fraud Prevention using Genetic Algorithm Solution,” which application is a continuation of U.S. application Ser. No. 11/593,962 now U.S. Pat. No. 7,657,497, filed Nov. 7, 2006, entitled “Online Fraud Prevention using Genetic Algorithm Solution,” which applications are incorporated herein by reference in their entirety.
The present application relates generally to the technical field of algorithms and programming and, in one specific example, to the use of genetic algorithms and programming to optimize fraud prevention rules as embodied in software as implemented in a computer.
The problem of online fraud is one that has challenged businesses engaging in electronic commerce (“E-commerce”) since the Internet was initially used as a means to buy and sell good and services. One aspect of the problem is that as the persons committing the fraud (fraudsters) tend to become more sophisticated with every new technological innovation designed to stop or deter them. More to the point, with every new anti-fraud innovation, the fraudsters respond with an even more sophisticated form of fraud. In short, a race exists whereby the fraudsters create a new type of fraudulent scheme, and the E-commerce business must develop a way to detect this fraud.
Many times the level of fraud committed by a fraudster is a function of the time that they have to commit the fraud. That is, the more time the fraudster has to commit the fraud, the greater fraud. This time aspect is, in turn, contingent upon the ability of the E-commerce business to discover the fraud such that the longer it takes the E-commerce business to discover the fraud, the longer the fraudster has to commit the fraud and, hence the greater the fraud.
Fraudulent schemes involving E-commerce can be as elementary as using stolen credit card numbers to purchase goods and services on line, to more sophisticated schemes whereby false email addresses, buyer or seller information is used to transact business. One of the most common types of fraud involves Internet auction sites. According to the Federal Trade Commission and Internet Fraud Watch, fraudulent schemes appearing on online auction sites are the most frequently reported form of Internet fraud. These schemes, and similar schemes for online retail goods, typically purport to offer high-value items—ranging from Cartier™ watches to computers to collectibles such as Beanie Babies™—that are likely to attract many consumers. These schemes induce their victims to send money for the promised items, but then deliver nothing or only an item far less valuable than what was promised (e.g., counterfeit or altered goods). While government policing agencies have stepped up efforts to thwart this fraud, fraudulent schemes still exist.
In response to these various fraudulent schemes, E-commerce sites have developed a number of types of tools to deal with these schemes. One type of tool developed by E-commerce sites is the use of various statistical models that use logistical regression, or some other type of statistical regression model, to look for and determine the long terms buying and selling trends of users, and using these trends, analyzing groups of online sales for fraud. For example, if transactions originating from a particular geographic location for a particular type of good or service are determined to commonly involve fraud (e.g., Cartier™ watches being sold by sellers located in Antarctica), then statistical modeling will typically be able to capture such an instance of fraud. Another tool used by various E-commerce sites is rule based modeling. In rules based modeling, specific rules are generated to combat and address specific fraudulent schemes. For example, if a particular seller from a particular geographical region (e.g., John Smith from California) is known to be a fraudster, then a rule can be set up to prevent this person from transacting business on a particular E-commerce site.
Both statistical models and rules based modeling have limitations in their abilities to detect fraud. For example, statistical models are very slow to detect fraud and suffer greatly from being unable to meet the ever increasing sophistication of fraudsters. Moreover, such models fail to segment well. That is, they rarely are able to provide specific information regarding specific instances of fraud; rather they typically provide a broad overview regarding losses from fraud generally. Using the above example, while statistical models may be able to tell you the amount of loss due to fraud in monetary terms, they typically would not be able to tell you the exact time, and date of a particular loss, not unless a further statistical model was generated to analyze time and date data.
A limitation for rule based modeling comes in the form of the reductionism that is inherent in rules based modeling. That is, the rules in rules based modeling are either applied or not applied to set of facts, and nuances within this set of facts is typically ignored. Again, using the above example, if a rule is created to flag every transaction by a person named “John Smith” who resides in California, then while the Fraudster named John Smith who resides in California might be stopped, so would all legitimate transactions by persons named John Smith who live in California. Other problems arising from rules based modeling come in the form of the fact that each rule must typically be manually entered into a system (e.g., someone must actually write and implement a rule that prohibits “John Smith from California” from engaging in transactions). This manual entering of rules has various computational costs associated with it. More to the point, many times, during the course of applying these manually entered rules, each rule must be applied to each set of facts such that as the number of rules grow, so does the computing times necessary to apply these rules. In some cases, this can result in instances where the computational complexity associated with applying these rules is on the order of O(n).
Not only can the computational costs associated with rules based modeling be costly, but the actual financial costs can also be costly. Many E-commerce sites have an elaborate, labor intensive, security apparatus for protecting against fraud. For example, when a fraudulent scheme is discovered, transactions associated with this scheme are, in many cases, flagged by the particular E-commerce site (e.g., eBay™) on which the transaction is taking place. Flagging a transaction means that the transaction will be examined more closely and, in some cases, stopped by the E-commerce site. In some instances, not only is the transaction flagged, but the account used in facilitating the fraud is also flagged. Many E-commerce sites have an entire fraud escalation apparatus, whereby potential fraud is identified and then the facts regarding the potential fraud are escalated to additional persons for further review and a determination regarding the existence of fraud. This process of identifying fraud and escalating fraud may involve two of more tiers of review, with a first tier of review by, for example, Floor Agents (or Regular Agents) reviewing data posted on the E-commerce by sellers of goods or services. Then, there is a second tier of review by Risk Agents (or Senior Agents and Risk Analysts) who review the flagged transactions and determine where the transaction is fraudulent. In those cases where the transaction is deemed fraudulent, yet no rule exists for this particular type of fraudulent scheme, then a new rule is generated by those at the second tier of review. In some cases, there is even a third tier of review by those Management and Tracking Agents charged with reviewing the activities of the Risk Agents.
In some cases, the phenomena of information loss occurs between the various tiers of review. For example, in some cases certain factors that go into the detection of fraud by those at the Regular Agent level cannot be easily or accurately conveyed to those at the higher tiers such as the Risk Agents and their managers. These factors can, among other things, include intuition.
BRIEF DESCRIPTION OF THE DRAWINGS
Some embodiments are illustrated by way of example and not limitation in the figures of the accompanying drawings in which:
Example methods and systems to enhance fraud detection are described. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of example embodiments. It will be evident, however, to one skilled in the art that the present invention may be practiced without these specific details.
A Three-Tier Architecture
In some embodiments, a set of potentially fraudulent transactions are examined by a Regular Agent 107 or other suitable person. The Regular Agent 107 looks for factual similarities amongst these transactions, similarities upon which a rule could be based, These similarities could include similar email information, bank financial information (e.g., a bank account number), or other similar information. Next a Regular Agent 107 generates a rule based upon these factual similarities using a GUI. During the course of generating this rule, the Regular Agent 107 may use plain text in the form of metadata to provide a plain text descriptions of the rule, including, for example, contextually driven information related to the rifle (e.g., goods purchased, buyer and seller information). This metadata may also include ID information for the particular fraud ID rule, ID information to be used in a fraud queue. Once the rule is created, is it parsed and placed into a data structure such as a Hash Table or other data structure with sub-linear performance (e.g., O(nlogn)) such as a Heap, Binary Search Tree, Radix Tree, Black-Red Tree or other suitable data structure. All the rules taken together are known as the fraud rule set 103. Once a fraud rule set 103 is created, this set can be used to process batched data that is warehoused over a particular period of time, and/or online/real time data. Some example embodiments may include the processing of data including applying the rules to the data. to determine whether the rules have been tripped or otherwise found to be true. Once a rule is found to be true, then the Rule ID and data in the form of an Account Number are stored together or separately in some type of persistent storage medium as, for example, a transaction type in a database table. Only the Rule ID information is stored.
Some example embodiments may include data being filtered via a filtering and priority module 106 working in combination with a Tracking Module. While this filtering process will be more fully discussed below, it suffices to say that a Hit Rate Table is generated to track the success of a rule over some predetermined period of time. This period of time can be predefined by a user, or can be determined automatically by the system. The rules that are not filtered are those rules and accounts which are flagged by these rules, for which there is some predetermined probability of fraud occurring based upon the rule being found to be true when applied to the account and facts underlying the account.
The GUI 200 may include a radio button 205 to allow a user to choose what setting they want, here a user can choose tracking mode or fraud queue (pictured but not referenced). Tracking allows a user to track the success of a particular rule, whereas the fraud queue setting allows a user to place the results of a particular rule into a fraud queue for analysis in addition to tracking the success of that rule. The GUI 200 may include a text box that allows a user to enter a suspect email address to match against the data found in a particular data set (e.g., match against an email address believe to be potentially fraudulent or involved in a fraudulent scheme).
The GUI 200 may also include a text box 207 to allow a user to enter in credit card information, which such information may be associated with a fraudulent account, of fraudulent scheme. A text box 208 allows a user to enter account data in the form of account type to match against a potentially fraudulent account. Example embodiments may further include a text box 209 into which a neural network score may be entered so as to allow a user to determine the applicability of a neural network model to be used to determine the existence of fraud. A spoof flag check box 212 is implemented to allow a user to flag an account as potentially spoofed (e.g., the subject of a fraudulent scheme). A text box 213 allows a potential rule to be compared against a particular fraud model.
An Interface Level
An example embodiment uses a client-based browser application, whereas other embodiments may be implemented via a command line interface. Some example embodiments may include an Application Programming Interface (API) implemented to allow one application to communicate with another. Some well-known client-based browser applications include Netscape Browsers™, Internet Explorer™, Mozilla Firefox™, Opera™, or some other suitable browser application. Common to these browser applications, is the ability to utilize a hyper-text transfer protocol (HTTP) or secured hyper-text transfer protocol (HTTPS) to get, upload (e.g., PUT) or delete web pages and interpret these web pages which are written in HTML and/or XML. HTTP and HTTPS are well known in the art, as are HTML and XML. (See XML for the World Wide Web, by Elizabeth Castro, Peachpit Press, 2000; Data on the Web: From Relations to Semistructured Data and XML 1ST Edition, by Serge Abiteboul, Peter Buneman, & Dan Suciu, Morgan Kaufmann, 1999.) HTTP and HTTPS are used in conjunction with a TCP/IP protocol as described in the OSI model, or the TCP protocol stack model, both of which are well known in the art. (See Computer Networking: A Top-Down Approach Featuring the Internet 2nd Edition, James F. Kurose and Keith W. Ross, Addision-Wesley, 2003.) The practical purpose of the client-based browser application is to enable a user to interact with the application through the display of plain text, and/or interactive, dynamic functionality in the form of buttons, text boxes, scroll down bars or other objects, widgets contained on one or more web pages constructed using the aforementioned HTML and/or XML.
Web pages are typically static or dynamic in nature. Those that are static typically display text as one would see it on a printed, physical page. Dynamic web pages, however, are interactive and allow for a user to input data, query data, and/or modify data just to name a few of the functionalities associated with dynamic web pages. The dynamic nature of web pages is a product of the use of the other technologies in combination with HTML and/or XML.
Some example embodiments may include, for example, a GUI used and implemented via a Java Servlet, Applet, or VBScript or C# form, or some other suitable programming language. The GUI may be written using the Perl programming language. As will be discussed below, web pages containing GUIs are stored at the Storage level, but executed at the Interface level via a web browser. These server pages contain objects such as text boxes, buttons, scroll-down bars, widgets, or some other suitable dynamic interface object. These objects, and the routines governing them, allow a user to retrieve, input, or delete content, just to name few of the functions. For example a user will be prompted with a new rule server page requesting the user to enter a new rule into a series of text boxes. These rules can be in the form of certain words associated together, words that make up information related to buyer or seller information on an E-commerce site. Once the new rule is entered, the rule will be sent to the Logic level, described below, where it will be stored for future use.
In some embodiments, the GUI 300 restricts the form by which a user can generate and enter rules into the system, such that the entries must reflect a particular grammar (see below). Some example embodiments may include the GUI allowing for a user to enter a rule in any logical form, and once entered these rules will be converted by the Logic level below into a form that can be processed by the particular grammar used by the system. This logical form will be a Boolean based algebra. Moreover the rules entered will be converted to the grammar used by the Logic level through performing various Boolean algebraic operations on the input.
Some example embodiments may include the above described Servlets, Applets, and/or VBScript forms being stored as server pages on one or more remote server computers connected to the client computer via an Internet. These remote servers can be a web server and/or application server. Web servers running JSP™ can include the Apache™/Apache Tomcat™ web server. Web servers running ASP™ can include a Microsoft Window Web Server 2003™ utilizing Internet Information Services (IIS). Application servers running JSP™ can include the Orion Application Server, or other J2EE™ certified application servers. Application servers running ASP™ can include Windows Server 2003™.
In some embodiments, the Logic level is governed by a scripting language that controls how and when certain web pages or pieces of content are provided to, or made accessible to, a particular user. This scripting language can be in the form of Java™, Perl, Python, or some other general purpose scripting language. For example, once the logic of a JSP determines that a particular object (e.g., a text box) on a web page has been executed (e.g., rules data has been entered and sent), the data from this text box is inputted, and sent to a web or application server. It is the routine written in a scripting language that determines whether, for example, the rules data is valid (e.g., that a properly constructed string has been entered). Some example embodiments may include the routine written in a scripting language to serve to retrieve data from a storage, data structure, or database level. The Storage level will be run by a separate database application, while, in other embodiments, a database embedded with a Logic level will be implemented.
In some example embodiments, a GUI implemented using the Perl programming language serves as an interface for a logic level written in an object oriented programming language such as C++, Java™, Delphi™, C#™ of some other suitable programming language. Some example embodiments may include a structured programming language such as C to implement the Logic level. This Perl implanted GUI may be located on a remote computer connected to another computer, in a client-server, peer-to-peer or some other suitable configuration. In some example embodiments, the routine written in, for example, C will retrieve data from a storage, data structure, or database level. Some example embodiments may include the Storage level being run by a separate database application, while in other embodiments a database embedded with a Logic level will be implemented. This separate database application will be located remotely on a separate computer, while, in other example embodiments, it will be located in the same computer.
In some embodiments, an Interface and Logic level is implemented to allow for the easy generation of various rules to be used in a Rule based modeling system used to detect fraud being committed on an E-commerce site. Some example embodiments may include the various tiers of review common to many fraud detection apparatus being replaced with fewer tiers, or even one tier of transaction review. In one embodiment, the Regular Agents 107 perform many of the same functions as Risk Agents, and more to the point, these Regular Agents 107 engage in actual rules generation (e.g., they create the rules for a rules based modeling System), as opposed to the Service Agents and Risk Analysts 101 generating such rules. Rules are generated for even one factual scenario (e.g., a single transaction) where the presence of fraud is believed to exist. In some example embodiments, rather than a testing bed being created and implemented by a Senior Agent and Risk Analysts 101 or others prior to the generation of the rules for a rules based modeling system, a Regular Agent 107 merely generates a rule and implements in the system. Some example embodiments may include rules to prevent fraud being generated on the fly, as needed by a Regular Agent 107 or other suitable person who directly examines transaction data for the a particular E-commerce site.
An Example Algorithm-Rules Engine
In some example embodiments, a rules engine 104 works together with a Tracking, Filtering and Priority Module and, in some embodiments, a Genetic Algorithm Module. Some example embodiments may include the rules engine 104 being used to create the rules and store the rules into some type of data structure. A method may be implemented including taking a stream of transaction data such as, for example, the combination of email information (e.g., “STOLENLAPTOP@email.com” or “FRAUDSELLER@email.com”) data and particular financial number data (e.g., “7562” or “6788”), and generating a rule based upon this data. For example, every time the email information “STOLENLAPTOP@email.com” and the financial number “7562” appear in combination together in a transaction, a rule is activated and the account or transaction is flagged as potentially fraudulent. Other types of data could be used in conjunction with, or in lieu, of the above described email information and financial number data. For example the following types of information could be used to generate one or more rules: first name, last name, email address, credit card number, type of credit card, bank account number, account activity, account status, known bad information (e.g., using a credit card marked as stolen or lost), country of residence, transaction details, transaction history, fraud model hits, neural network scores, or some other suitable type of data. Some example embodiments may include the following neural networks: Kohonen self-organizing network, Recurrent networks, Simple recurrent networks, Hopfield networks, Stochastic neural networks, Boltzmann machines, Modular neural networks, Committee of machines, Associative Neural Network (ASNN), Holographic associative memory, Instantaneously trained networks, Spiking neural networks, Dynamic neural networks, Cascading neural networks, Neuro-fuzzy networks, or some other suitable neural network. In some example embodiments, a Hidden Markov model is used to generate data.
In some embodiments, the rules engine 104 operates based upon certain principles of compiler design and implementation as are known in the art. (See Compilers: Principles, Techniques, and Tools, by Alfred V. Aho, Ravi Sethi, Jeffrey D. Ullman, Addison Wesley, 1986.) Some example embodiments may include a grammar created with the Backus-Naur form (BNF) of:
This grammar is an unrestricted grammar, while in other embodiments it is a context sensitive, context free, or regular grammar as is known in the art. (See An Introduction to Formal Language and Automata, by Peter Linz, Jones and Bartlett, 2001.) This grammar has complex expressions that allow for “ANDs” of “OR”, but which do not allow for “ORs” of “ANDs”. As noted above, rules are not limited to the types of data described in the above grammar (e.g., email information and financial information), but can be broader.
Some example embodiments may include this grammar being used to parse input in the form of user generated or machine generated (see discussion below relating to the genetic algorithm module 114) rules (e.g., a rules set) into one or more parse trees or some other suitable data structure with sub-linear performance. These rules are received via a GUI, whereas, in other embodiments, these rules are received via an API. These parse trees have the following characteristics: the leaf nodes are a rule, the leaf nodes are combined with a Boolean operator (e.g., “AND” or “OR”) to form a subtree, the tree has a maximum depth of two (2) nodes, a tree can have an unlimited number of children and is not limited to being a Binary Search Tree, every subtree is internally considered to be a tree, every subtree has a rule of the grammar associated with it, there can be more than one parse tree implemented by, for example, the rules engine 104.
Some example embodiments may include each of these parse trees being evaluated in a bottom up manner starting with the leaf nodes. Specifically, for every subtree in the parse tree one or more data structures with sub-linear performance is created to store one more nodes of this subtree. In some cases, a Hash Table is used to store the parse tree and its various subtrees, whereas, in other embodiments, a Radix Tree, Binary Search Tree, Heap, or Red-Black Tree is implemented in lieu of, or, in combination with, the Hash Table. When viewed in terms of computational efficiency a Hash Table may be used instead of, for example, a Radix Tree, whereas, in other embodiments, the Radix Tree is, for example, may be used. This issue of which data structure to use may be based on an appreciation that while a Hash Table has amortized O(1) performance, for certain inputs (e.g., data sets) a Hash Table has O(n) performance. In such cases, a data structure such as, for example, a Radix Tree with actual O(21gn) worse case performance may be used instead of a Hash Table.
In some embodiments, once each parse tree is stored into some type of data structure, such as a Hash Table, then the rule is applied to various transaction data. This transaction data may be in the form of warehoused data, or may be in the form of data occurring real time, such as, online data. In the case of warehoused data individual accounts may be evaluated to determine whether a particular account is deemed to be fraudulent based upon the application of a rules set. Some example embodiments may include individual transactions being evaluated to determine whether they are fraudulent by using the rules. Where a rule is evaluated as true (e.g., the account has factual characteristics supporting the application of the rule), the applicable rule is mapped (e.g., mapped by a Rule ID value) to the account number, or transaction number. This mapping this then stored into a persistent storage medium (see e.g., below description of Storage level) for future use.
An Example Algorithm-Tracking Module
Some example embodiments may include implementing tracking functionality with the following characteristics: tracking is based upon the system as whole, the tracking interface allows for rule killing, the tracking module gives “credit” to a particular rule based upon the number of times a particular rule successfully flags a particular account as fraudulent. This functionality has the additional characteristics of tracking every subtree (e.g., rule) performance, including those rules not ultimately applied to flag an account or transaction as fraudulent. Additionally, this functionality allows for a user to test a particular rule, and it ability to detect a fraudulent account or transaction, without actually using the rule to flag or segregate out specific transactions. The hit rule module 111, or credit & allocation module 116 may implementing this functionality.
Some embodiments may include, a hit rule module 111 that contains a “Rules History Table” that allows for a user to compute the hit rate or success of a rule in flagging an account or transaction as fraudulent. The following is an example of such as table:
A “Hit Rate Table” is generated, in part, from a “Rules History Table”. Some characteristics of this “Hit Rate Table” include, for example, as alluded to above, rules that may be credited with detecting a fraudulent account or transaction, even if the rule is not given credit for restricting the account. The fact that credit is given, even in instances where the rule is not used as the primary basis for flagging a transaction, rewards users who write rules. The hit may be based upon a number or series of days, as opposed to the hit rate for a specific day. More to the point, given the ongoing nature of fraud, and the difficulty of discovery, credit for the discovery of the fraud is spread out over some specific time period, and a success rate attributed to the rule to reflect this time period. For example, if over a window of five (5) days, a rule detected one, and only one, account to be fraudulent for every single day, the rule may be credited for the last 3 days with a resulting hit rate of 60%. The following is a description of the “Hit Rate Table”:
In some embodiments, the tracking module facilitates an open review paradigm, wherein it allows a user to review the relative success of all other users, wherein one, but not the only, measure of success is the whether or not the rules that a user has created have been successful in flagging or restricting fraudulent accounts or transaction. This may include users who do create successful rules being identified by the system for reward. This system creates a very egalitarian system of performance review, wherein success of a user is easily observed.
An Example Algorithm-Filtering and Priority Engine
In some embodiments, filtering and priority functionality is implemented with the following characteristics: it may use the “Hit Rate Table” to filter successful rules from unsuccessful rules, it may, filter based upon metadata descriptions supplied by a user at the time a rule is created, it may or may not filter based upon specific requests contained in a particular rule's metadata (e.g., a rule may contain a metadata tag that requests that the rule always be used and never filtered), it may de-duplicate the priority queue or other data structure used to store the flagged or restrict account or transaction such that any flagged account or transaction will only be flagged by one rule (e.g., that rule that is most successful in detecting the fraud present in the account or transaction), it may sort the rules in order of priority such that account with the greatest probability of fraud will be at the top of the queue. Some example embodiments may include the filter & priority module 106 and reaper 113 being used to implement this functionality.
Some embodiments may include, a reaper 113 that removes ineffective rules (e.g., rules that have shown a history of being unable to detect fraudulent accounts or transactions). The reaper 113 merely prevents an unsuccessful rule from being used by the system for the actual flagging of an account or transaction. Despite this, such a rule can accrue a value in the “Rules History Table” or the “Hit Rate” table, but will not be allowed serve as the basis for restricting an account. The purpose of still allowing an ineffective rule to be used for the purpose of accruing, for example, a hit rate is so that when a fraud scheme arises that the previously ineffective rule could detect, then the hypothetical success of the rule could be determined and the rule can be re-activated for use.
In some embodiments, once an account or transaction is placed into the queue, it is considered ready for human review. One example embodiments may include more than one queue being used such that queues, and the transactions that they contain, can be segregated based upon the expertise of the human reviewer. For example, flagged transactions arising from Antarctica, can be placed into a queue for review by people with an expertise in Antarctica. And again, flagged accounts arising from North Korea, can be placed into a queue for review by people with an expertise in speaking or who are knowledgeable about North Koreans.
An Example Algorithm-Genetic Programming Algorithm
Some example embodiments may include a code module (e.g., genetic algorithm module 114) implementing principles of Genetic programming being employed to generate the most “fit” rules based upon the goals of a particular user. This code module may use a library written in the C programming language called Lilgp. In general, Genetic programming involves concepts of: initialization, selection, reproduction and ultimately termination. As with many biological based systems there is a notion of natural selection and fitness with only the most “fit” species in a particular biological system surviving. It is through this same notion of fitness, applied in the present context that only the most “fit” rules are allowed to survive. The term “fit” may include the following definition:
- Let O be the set of all rules combined together into islands of trees (e.g., a set of trees) via various Boolean operations;
- Let R be the set of all restricted accounts (e.g., accounts for which fraud is believed to exist) for a particular time period;
- Let Ai be the set of accounts found by the ith tree in O
- Then take (|Ai∩R|)/|Ai|=as hit rate for the ith tree in O.
The term “fit” can be described by a second definition of fitness that takes into account certain economic considerations. For example, the notion of a “fit” rule can be based upon the return on investment generate by the rule (ROI). For example:
- Let O be the set of all rules combined together into islands of trees (e.g., a set of trees) via various Boolean operations;
- Let R be the set of all restricted accounts (e.g., accounts for which fraud is believed to exist) for a particular time period;
- Let Ai be the set of accounts found by the ith tree in O
- Let λ be the minimum expected savings in finding a restricted account;
- Then take (|Ai∩R|)*λ−|Ai|=a return on investment amount from the ith tree in O.
Clearly, ROI is not the only criterion (economic or otherwise) that could be used to define whether a function is “fit”. Other criterion such as overall savings, saving for a particular time period, or some other suitable economic criterion may also be used to describe a “fit” rule.
In addition to the definition of whether a rule is “fit”, there are additional concepts that are defined in some embodiments of the present system. For example there is a concept of “crossover” and one of “mutation”. These concepts generally relate to the above referenced concept of reproduction. As will be more fully described below, “crossover” involves taking two trees from the same or different islands of trees, randomly picking a leaf node (e.g., a rule) or a subtree, and swapping this subtree with a randomly picked subtree from a different tree on the same or different island of trees. Randomness may be achieved using a pseudo-random number generator. In one further embodiment, “crossover” results in two new trees are generated for every two parents. A tree can be a single leaf (e.g., a rule).
In some embodiments, a rule set is received from, for example, a user, Genetic programming algorithm, a group of users using the rule set authored by them over a time period or other suitable source. This rule set is mapped to account data, and the set of account numbers flagged by that rule are stored into some type of persistent or non-persistent data structure. These accounts then constitute a target set (e.g., fraudulent accounts for the particular fraud that sought to be detected). A population is then created by generating rule trees by taking random logical combination (e.g., unions and intersections) of the rules from the initial rules set. For each rule tree in the population, the rule tree is executed (e.g. applied to a data set including, for example, account information) to determine the set of accounts captured by the rule tree (e.g., stepping through the tree taking the union and the intersections of component rules). Moreover, for each rule tree in the population, the fitness of the rule tree is computed by taking the intersection of the set of accounts determined from the set of accounts captured by the tree (see above) and the set of accounts used as the target. The resulting rule trees are ranked according to their fitness. Furthermore, a new population of trees is created by applying cross-over and mutation to the top ranking trees. The notion of what constitutes a top ranking tree is based upon some predetermined percentage value such as, for example, the top 10%, 20%, 30%, 40%, or even the top 50%. This value can be manually determined by a user, or automatically determined by a computer system. If a termination condition is met, then the resulting new fit tree or set of trees is injected back into a rules engine such as, for example, rules engine 104. A termination condition can occur after a set number of iterations or migration as referenced elsewhere herein. Where a termination condition is not met, then the step of stepping through the tree taking the union and the intersections of component rules is repeated, as are the steps subsequent to this step.
In some embodiments, as with other natural selection based systems, the Genetic programming as employed has certain additional characteristics. For example the most “fit” tree survives and is not subjected to “mutation”, or “crossover”. This is known as “elitism”. Moreover, some example embodiments may include a new set of child trees being generated through “crossover”, “mutation,” and these trees are used to replace the most unfit trees. So if, for example, “N” new trees reproduced via crossover, “N” unfit trees are replaced. With regard to the “mutation”, only a defined percentage of trees are ever mutated.
Applying these various concepts in the present context, in some embodiments, a user (e.g., a Regular Agent 107) will choose a particular rule, apply the concepts of “crossover” and mutation” to discover fit rules that describe a set of accounts that most closely match the set of fraudulent accounts in the chosen rule, and inject the new fit rules back into the system for use by the system in combination with, or separately from, various human created rules. Some example embodiments may include the initial choice or a particular rule being based upon the goals that the user is attempting effectuate. For example, if the user is looking to generate rules relating to accounts held by people from Antarctica, then the user will create a rule or select a rule such as “all accounts held by people from Antarctica”. The system will then find the intersection between this rule, and all restricted accounts (e.g., accounts where fraud was found to exist). Logically, this can be presented in the following manner:
- Let L be the rule namely the set of “all accounts held by people from Antarctica”;
- Let R be the set of all account determined to be restricted;
- L∩R=R′, where R′ is the subset of all restricted/bad accounts that are held by people from Antarctica.
Once this subset R′ is created, the GP module (e.g., genetic algorithm module 114) will proceed to discover the most effective or fit rules that best capture the data in R′ by using the process of
FIG. 8using R′ as the target set in 802. These new rules will be discovered through the iterative application of the processes of crossover and mutation to the fit members of the population. Once these various processes are applied a plurality of times the fit new rules are injected back into the system. Some example embodiments may include “crossover” and “mutation” occurring over a number of generations, iterations. For example the “crossover” and “mutation” process can be repeated 100, 200, 300, 400, 500, 600, or even 1000 times prior to being injected back into the system. Moreover, as with natural systems, there is a concept of “migration” described and implemented. Specifically, after every so many iterations, trees and subtrees from one set or island is allowed to reproduce (e.g., “crossover” and/or “mutate”) with trees and subtrees from another set or island. For example, in some embodiments, after every 20, 30, 40, 50, 60, 70, or even 100 iterations “migration” is allowed between sets or islands of rules. Injecting these rules back into the system allow for experimentation with new rules at a pace faster than these rules can be created by humans, or more conventional means (e.g., the generation of rules and use of these rules in various testing scenarios, test beds or the like). Once the reproduced rules are created, the user receives a file containing all the new reproduced rules combined into a tree form with various Boolean operators. The user, at this point, may decide whether or not to manually inject these rules back into the system.
Some embodiments may include a Storage level that is implemented whereby tables of data are created, and data is inserted into, selected from, these tables using a structured query language (SQL) or some other database-related language known in the art. (See The Fundamentals of Database Systems 3rd Edition, by Remez Elmasri & Shamkant B. Navathe, Addison-Wesley, 2000.) These tables of data can be managed using a database application such as, for example, MySQL™, SQLServer™, Oracle 8i™ or 10g™, or some other suitable database application. These tables are organized into a relational-database schema (RDS) or object-relational-database schemas (ORDS), as is known in the art. (See Id.) These schemas can be normalized using certain normalization algorithms so as to avoid abnormalities such as non-additive joins and other problems. Additionally, these normalization algorithms include Boyce-Codd Normal Form or some other normalization, optimization algorithm known in the art. (See Id.)
In some embodiments, various types of tree data structures are used to organize data for the purposed of searching. These tree data structures include Radix Trees, Red-black trees, Binary Search Tree, Heaps, or some other suitable data structure as is known in the art. (See Algorithms in C++3rd Edition: Parts 1-4, Robert Sedgewick, Addison-Wesley, 1998.) Some other type of data structure may be used such as a One or Multi-Dimensional Array, Hash Table, Link List, Queue, Stack, or some other suitable data structure as is known in the art. (See Id.) Once implemented, these data structures can then be searched using a divide and conquer, or other search methods well known in the art. (See Id.)
Some example embodiments may include the above described three (3) tiers or levels being written as one or more a software modules with each module contributing to the functionality of each level or tier. Common too many of these modules is the ability to generate, use and manipulate the above described data and data sets. These modules, and associated functionality, may be used by either the client, server, or peer applications. These various modules can be implemented into the system on an as-needed basis. These modules may be written in an object-oriented-computer language such that a component oriented or object-oriented programming technique can be implemented using, a Visual Component Library (VCL), Component Library for Cross Platform (CLX), Java Beans (JB), Java Enterprise Beans (EJB), Component Object Model (COM), or Distributed Component Object Model (DCOM) or other suitable technique. These modules are linked to other modules via various APIs and then compiled into one complete server and/or client application. The process for using modules in the building of client and server applications is well known in the art. (See Component Based Software Engineering: Putting the Pieces Together, by George T. Heineman and William T. Council, Addison-Wesley, 2001; Delphi Component Design, by Danny Thorpe, Addison-Wesley, 1996.) Further, these modules, and the tiers that they make up, are linked together via various distributed programming protocols as distributed computing modules.
Distributed Computing Modules
Some example embodiments may include remote procedure calls being used to implement one or more of the above described levels of the three-tier architecture across a distributed programming environment. For example, a Logic level resides on a first computer system that is remotely located from a second computer system containing an Interface or Storage level. These first and second computer systems can be configured in a server-client, peer-to-peer or some other configuration. These various levels can be written using the above described component design principles, and can be written in the same programming language, or a different programming language. Various protocols are implemented, to enable these various levels, and components contained therein, to communicate regardless of the programming language used to write these components. For example, a module written in C++ using COBRA or SOAP can communicate with another remote module written in Java™. These protocols include Simple Object Access Protocol (SOAP), and the Common Object Request Broker Architecture (CORBA) or some other suitable protocol. These protocols are well-known in the art. (See The CORBA Reference Guide: Understanding the Common Object Request Broker Architecture, by Alan Pope, Addison-Wesley, 1997.)
A System of Transmission Between a Server and Client
In some embodiments, the above described components that make up the platform architecture communicate using the OSI or TCP/IP protocol stack models for defining network protocols that facilitate the transmission of data. Applying these models, a system of data transmission between a server and client computer system can be described as a series of roughly five layers comprising as a: physical layer, data link layer, network layer, transport layer and application layer. (See Open System Networking: TCP/IP and OSI, by David M. Piscitello & A. Lyman Chapin, Addison-Wesley, 1993.) Some example embodiments may include the various levels (e.g., the Interface, Logic and Storage levels) residing on the application layer of the TCP/IP protocol stack. The present application may utilize HTTP to transmit content between the server and client applications, whereas in other embodiments another protocol known in the art is utilized. Content from an application residing at the application layer is loaded into the data load field of a TCP segment residing at the transport layer. This TCP segment also contains port information for a recipient application a module residing remotely. This TCP segment is loaded into the data field of an IP or UDP datagram residing at the network layer. Next, this IP datagram is loaded into a frame residing at the data link layer. This frame is then encoded at the physical layer and the content transmitted over a network such as an Internet, local area network (LAN) or wide area network (WAN). The terms Internet refers to a network of networks. Such networks may use a variety of protocols for exchange of information, such as TCP/IP, ATM, SNA, SDI, etc, and may be used within a variety of topologies or structures. This network may include a code sensing multiple access network (CSMA) such an Ethernet based network. This network may include a code divisional multiple access (CDMA) network, or some other suitable network.
In some embodiments, a digital processing system or computer system is implemented that includes a processor, which may represent one or more processors and may include one or more conventional types of such processors (e.g., x86, x86-64), such as an AMD processor, Intel Pentium processor or other suitable processor. A memory is coupled to the processor by a bus. The memory may be a dynamic random access memory (DRAM) and/or may include static RAM (SRAM). The processor may also be coupled to other types of storage areas/memories (e.g., cache, Flash memory, disk, etc.), which could be considered as part of the memory or separate from the memory.
Some example embodiments may include a bus further coupling the processor to a display controller, a mass memory or some type of computer-readable medium device, a modem or network interface card or adaptor, and an input/output (I/O) controller. The display controller controls, in a conventional manner, a display, which may represent a cathode ray tube (CRT) display, a liquid crystal display (LCD), a plasma display, or other type of suitable display device. Computer-readable medium may include a mass memory magnetic, optical, magneto-optical, tape, and/or other type of machine-readable medium/device for storing information. For example, the computer-readable medium may represent a hard disk, a read-only or writeable optical CD, etc. Some example embodiments may include a network adaptor card such as a modem or network interface card being used to exchange data across a network such as an Internet. The I/O controller controls I/O device(s), which may include one or more keyboards, mouse/trackball, or other pointing devices, magnetic and/or optical disk drives, printers, scanners, digital cameras, microphones, etc.
In some embodiments, an embodiment may be implemented entirely in executable computer program instructions which are stored on a computer-readable medium or may be implemented in a combination of software and hardware, or entirely in hardware via circuits such as logic circuits.
Embodiments within the scope of the present invention include computer-readable medium for carrying or having computer-executable instructions or data structures stored thereon. Such computer-readable medium may be any available medium, which is accessible by a general-purpose or special-purpose computer system. By way of example, and not limitation, such computer-readable medium can comprise physical storage medium such as RAM, ROM, EPROM, CD-ROM, or other optical-disk storage, magnetic-disk storage or other magnetic-storage devices, or any other medium which can be used to carry or store desired program code means in the form of computer-executable instructions, computer-readable instructions, or data structures and which may be accessed by a general-purpose or special-purpose computer system. This physical storage medium may be fixed to the computer system as in the case of a magnetic drive or removable as in the case of an EEPROM device (e.g., flash memory device).
In some embodiments, when information is transferred or provided over a network or another communications connection (e.g., either hardwired, wireless, or a combination of hardwired or wireless) to a computer system, the connection is properly viewed as a computer-readable medium. Thus, any such connection is properly termed a computer-readable medium. Combinations of the above should also be included within the scope of computer-readable medium. Computer-executable or computer-readable instructions comprise, for example, instructions and data which cause a general-purpose computer system or special-purpose computer system to perform a certain function or group of functions. The computer-executable or computer-readable instructions may be, for example, binaries, or intermediate format instructions such as assembly language, or even source code.
In this description, and in the following claims, a computer system is defined as one or more software modules, one or more hardware modules, or combinations thereof, that work together to perform operations on electronic data. For example, the definition of computer system includes the hardware modules of a personal computer, as well as software modules, such as the operating system of the personal computer. The physical layout of the modules is not important. A computer system may include one or more computers coupled via a network. Likewise, a computer system may include a single physical device (e.g., a mobile phone or Personal Digital Assistant (PDA)) where internal modules (e.g., a processor and memory) work together to perform operations on electronic data.
Some embodiments may be practiced in network computing environments with many types of computer system configurations, including hubs, routers, wireless access points (APs), wireless stations, personal computers, laptop computers, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, mobile telephones, PDAs, pagers, and the like. One embodiment can also be practiced in distributed system environments where local and remote computer systems, which are linked (e.g., either by hardwired, wireless, or a combination of hardwired and wireless connections) through a network, both perform tasks. In a distributed system environment, program modules may be located in both local and remote memory-storage devices (see below).
The example computer system 1500 includes a processor 1502 (e.g., a central processing unit (CPU), a graphics processing unit (GPU) or both), a main memory 1501 and a static memory 1506, which communicate with each other via a bus 1508. The computer system 1500 may further include a video display unit 1510 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)). The computer system 1500 also includes an alphanumeric input device 1512 (e.g., a keyboard), a user interface (UI) cursor controller 1511 (e.g., a mouse), a disk drive unit 1516, a signal generation device 1518 (e.g., a speaker) and a network interface device (e.g., a transmitter) 1520.
The disk drive unit 1516 includes a machine-readable medium 1522 on which is stored one or more sets of instructions and data structures (e.g., software) embodying or utilized by any one or more of the methodologies or functions described herein. The software may also reside, completely or at least partially, within the main memory 1501 and/or within the processor 1502 during execution thereof by the computer system 1500, the main memory 1501 and the processor 1502 also constituting machine-readable media.
The instructions 1521 may further be transmitted or received over a network 1526 via the network interface device 1520 utilizing any one of a number of well-known transfer protocols (e.g., HTTP, SIP).
While the removable physical storage medium 101 is shown in an example embodiment to be a single medium, the term “machine-readable medium” should be taken to include a single medium or multiple medium (e.g., a centralized or distributed database, and/or associated caches and servers) that store the one or more sets of instructions. The term “machine-readable medium” shall also be taken to include any medium that is capable of storing, encoding or carrying a set of instructions for execution by the machine and that cause the machine to perform any of the one or more of the methodologies described herein. The term “machine-readable medium” shall accordingly be taken to include, but not be limited to, solid-state memories, optical and magnetic medium, and carrier wave signals.
Some embodiments, described herein, may have a variety of marketplace applications. For example no test bed is needed to generate rules. And again, some example embodiments may include on the fly rules generation and testing being performed. And again massive independent peer review of rules can take place. An API can be used by third-party vendors to, for example, create new rules directly, or by using the genetic algorithm module 114, or to use the rules engine 104 to actually test these created rules.
Some example embodiments may include no test bed being needed for implementing the various rules of the present invention. One upshot of this, with regard to marketplace applications, is that rather than having a highly organized testing structure, wherein test beds are generated and data is tested using these test beds, data and rules can be tested without resorting to some highly structured system of testing.
In some embodiments, on the fly rules generation in testing is implemented. The advantage of on the fly rules generation in testing is that rather than having an instance of fraud escalated to some type of supervisor who would, in turn, generate antifraud rules that can be generated by those who have first hand knowledge of the fraud taking place. These persons include those who have either examined the facts surrounding the fraud, or who have an intuition regarding the possible existence of the fraud. This allows individuals with this knowledge to generate rules and to test these rules without having to resort to supervisors who may be removed from the facts surrounding a particular instance of fraud, or removed from the intuition required to detect this fraud.
In some embodiments, massive independent peer review is facilitated. For example rules are open for review by anyone who is able to use the system. This allows for anyone to comment on the efficacy and applicability of these rules, and allows anyone to test these rules. In some embodiment, this is an advantageous over other systems which merely allow for only persons in a supervisory capacity to create, test and review rules. In this way the present system is very egalitarian in that it allows persons with first hand knowledge of a fraud to be able to generate rules to counter this fraud, rather than having individuals who do not have first hand knowledge of particular instance of fraud, but who are changed with generating rules based merely upon their employment position and seniority, to generate rules. By allowing large scale risk free experimentation by many individuals, some embodiments may allow bottom-up innovation to emerge in large organizations which is critical in a fast changing world such as fraud.
Some example embodiments may include further marketplace applications in the form of allowing third-parties to use various APIs as associated with the genetic algorithm module 114 and rules engine 104 to create new rules to be tested and modified. For example a particular third-party may have been subjected to a fraudulent scheme that a particular user, of one embodiment of the present invention, may not have been subjected to. The system may be rich and robust in terms of its ability to meet potential fraud.
In some example embodiments, users of the present system will be able to, for example, charge a fee to allow third parties to use the API as associated with, for example, the genetic algorithm module 114 or the rules engine 104. For example, a third party may want to test, and benefit from the creation of a rules set, through the use of the genetic algorithm module 114. More to the point, the genetic algorithm module 114 can be used to modify an existing rule set and, once the existing rule set is modified, the rules engine 104 can be used to test this rule set. A fee can be charged to allow the third parties to have access to this genetic algorithm module 114 API, and rules engine 104 API. (See e.g.,
Some example embodiments may include a further marketplace application in instances where there is a need to reduce labor costs associated with fraud prevention. For example labor costs can be reduced by creating a more linear or flat organizational structure within the particular fraud detection regime of a particular corporation. For example, rather than having a hierarchy of first-tier fraud agents, a second-tier of supervisory agents, and a third-tier of supervisory managers controlling or directing the actions of these other two-tiers one embodiment can allow for a single-tier to be used. Some example embodiments may include one implementation allowing for a corporation or company engaging in E-ecommerce to have a single-tier of persons reviewing fraudulent or potentially fraudulent transactions. More to the point the present system through, for example, its on the fly rules generation and testing, allows persons with first hand knowledge of fraudulent transactions (e.g., fraud agents alone) to generate as many rules as they would like irrespective of the costs associated with these rules. More to the point, the ability to generate a rule, using one embodiment of the present invention, is limited only by the time that it takes an agent to actually generate the rule. That is, there is little cost associated with escalating a potentially fraudulent account to a second or even third tier of review.
Some embodiments may include a system including: a receiver operatively coupled to one or more application servers for receiving a rules set whose function is to detect fraud, one or more database servers operatively coupled to the one or more application servers to facilitate mapping the rules set to a data set, and for mapping success data to each members of the rules set, wherein the one or more database servers is operatively coupled to the one or more application servers so as to facilitate the filtering of the members of the rule set, and wherein the one or more database servers is operatively coupled to the one or more application servers so as to facilitate the ordering of the members of the data set by giving priority to those members of the data set with the greatest probability for being fraudulent based upon the success data of each member of the rule set in detecting fraud.
In some example embodiments, a system is described as including: a receiver operatively coupled to one or more application servers for receiving a rules set whose function is to detect fraud, one or more database servers operatively coupled to the one or more application servers so as to facilitate the finding of the logical intersection of a set of accounts captured by each rule and a set containing data that has been determined to be fraudulent wherein the one or more application servers facilitates producing new rule members using the fit rule members from the logical intersection, and wherein the one or more application servers facilitates injecting the new rule members and the fit rule members into a Rules Engine.
Some example embodiments may include a method being described as including: receiving a rules set whose function is to detect fraud, mapping the rules set to a data set, mapping success data to each members of the rules set, filtering the members of the rule set, and ordering members of the data set by giving priority to those members of the data set with the greatest probability for being fraudulent based upon the success data of each member of the rule set in detecting fraud. The method is further described as including receiving the rules set via a GUI. The method is further described as including receiving the rules set via an API. The method also includes account information that makes up members of the data set. The methods also include storing the mapping of the rules set to the data set in a storage medium. Furthermore, the method additionally includes the mapping of the success data to each member of the rules set in a storage medium. Moreover, the method also includes filtering and removing a member of the rules set based upon metadata. Additionally, the method includes filtering wherein a member of the rules set is removed based upon the low probability of the member rule in being able to detect fraud. The method also includes filtering such that duplicate mappings of the members of the rules set to the members of the data set are removed, where a first member rule of the rule set has a higher success data value as compared to a second member rule of the rule set. The method also includes taking members of the data set and ordering them using a priority queue. This method is implemented in a machine-readable medium containing a set of instructions that are executable by a suitability programmed computer.
Some example embodiments may include a method being implemented that includes: receiving a rules set whose function is to detect fraud, finding the logical intersection of a set of accounts captured by each rule and a set containing data that has been determined to be fraudulent, selecting fit rule members from the logical intersection, producing new rule members using the fit rule members from the logical intersection, and injecting the new rule members and the fit rule members into a Rules Engine. Moreover the method further includes receiving the rules set via a GUI. Additionally, the method further includes receiving the rules set via an API. Furthermore the method further includes determining the fitness of the rule member based upon the return on investment for using the rule. The method may also include using crossover in producing a new member rule. The method also includes using mutation to produce a new member rule. The method may also include using a machine-readable medium containing a set of instructions that are executable by a suitability programmed computer to perform the method.
It is to be understood that the above description is intended to be illustrative, and not restrictive. Although numerous characteristics and advantages of various embodiments as described herein have been set forth in the foregoing description, together with details of the structure and function of various embodiments, many other embodiments and changes to details will be apparent to those of skill in the art upon reviewing the above description. The scope of the invention should be, therefore, determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled. In the appended claims, the terms “including” and “in which” are used as the plain-English equivalents of the respective terms “comprising” and “wherein,” respectively. Moreover, the terms “first,” “second,” and “third,” etc., are used merely as labels, and are not intended to impose numerical requirements on their objects.
The Abstract of the Disclosure is provided to comply with 37 C.F.R. §1.72(b), requiring an abstract that will allow the reader to quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. In addition, in the foregoing Detailed Description, it can be seen that various features are grouped together in a single embodiment for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Thus the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as a separate embodiment.
1. A system comprising:
- a receiver to receive a set of rules for use in detecting fraudulent transactions;
- a server, operatively connected to the receiver, the server configured to identify a set of accounts that facilitated the detected fraudulent transactions and further configured to generate a new rule set based on an intersection of at least two of the respective rules, the new rule set to identify fraudulent accounts based on a series of transactions and the intersection including at least one of crossover or mutation of the respective rules.
2. The system of claim 1, wherein the server is further to generate a tree-form based on the new rule set.
3. The system of claim 2, wherein the crossover is to randomly select a first node on a first tree-form and a second node on a second tree-form, and replace the first node with the second node and mutation is to randomly select a node and replace it and its children with a randomly selected leaf.
4. The system of claim 1, wherein the server is further to generate a set of hash tables based on the new rule set.
5. The system of claim 1, wherein the server is further to calculate a success rate of the new rule set in identifying the fraudulent accounts.
6. The system of claim 5, wherein the server is further to remove the new rule set if the success rate is below a predefined threshold.
7. The system of claim 6, wherein the removed new rule set is used for the purpose of accruing a hit rate.
8. The system of claim 7, wherein the server is further to re-activate the removed new rule set based on a subsequent hit rate.
9. The system of claim 1, wherein the new rule set is generated to capture data in a predefined set of known fraudulent transactions.
10. The system of claim 1, wherein the server is further to order the fraudulent accounts according to a calculated probability for fraud.
11. A method comprising:
- receiving a set of rules for use in detecting fraudulent transactions;
- identifying a set of accounts that facilitated the detected fraudulent transactions; and
- generating, using one or more processors, a new rule set based on an intersection of at least two of the respective rules, the new rule set to identify fraudulent accounts based on a series of transactions and the intersection including at least one of crossover or mutation of the respective rules.
12. The method of claim 11, further comprising generating a tree-form based on the new rule set.
13. The method of claim 11, further comprising generating a set of hash tables based on the new rule set.
14. The method of claim 11, further comprising calculating a success rate of the new rule set in identifying the fraudulent accounts.
15. The system of claim 14, further comprising removing the new rule set if the success rate is below a predefined threshold.
16. The system of claim 15, wherein the removed new rule set is used for the purpose of accruing a hit rate.
17. The system of claim 16, further comprising re-activating the removed new rule set based on a subsequent hit rate.
18. The method of claim 11, capturing data in a predefined set of known fraudulent transactions to generate the new rule set.
19. The method of claim 11, further comprising ordering the fraudulent accounts according to a calculated probability for fraud.
20. A non-transitory computer-readable storage medium having instructions embodied thereon, the instructions executable by a processor to perform a method of fraud detection, the method comprising:
- receiving a set of rules for use in detecting fraudulent transactions;
- identifying a set of accounts that facilitated the detected fraudulent transactions; and
- generating a new rule set based on an intersection of at least two of the respective rules, the new rule set to identify fraudulent accounts based on a series of transactions and the intersection including at least one of the crossover or mutation of the respective rules.
U.S. Patent Documents
|5822741||October 13, 1998||Fischthal|
|6029144||February 22, 2000||Barrett et al.|
|6094643||July 25, 2000||Anderson et al.|
|6157707||December 5, 2000||Baulier et al.|
|6449611||September 10, 2002||Frankel et al.|
|6601048||July 29, 2003||Gavan et al.|
|6850606||February 1, 2005||Lawyer et al.|
|6886008||April 26, 2005||Blayvas et al.|
|6941285||September 6, 2005||Sarcanin|
|6941287||September 6, 2005||Vaidyanathan et al.|
|7014108||March 21, 2006||Sorenson et al.|
|7089592||August 8, 2006||Adjaoute|
|7092398||August 15, 2006||Schweitzer|
|7093130||August 15, 2006||Kobayashi et al.|
|7136841||November 14, 2006||Cook|
|7263506||August 28, 2007||Lee et al.|
|7331521||February 19, 2008||Sorenson et al.|
|7376431||May 20, 2008||Niedermeyer|
|7403922||July 22, 2008||Lewis et al.|
|7433855||October 7, 2008||Gavan et al.|
|7436989||October 14, 2008||Chung et al.|
|7457401||November 25, 2008||Lawyer et al.|
|7587367||September 8, 2009||Mengerink|
|7657497||February 2, 2010||Nandy|
|7840520||November 23, 2010||Nandy|
|8019678||September 13, 2011||Wright et al.|
|20020161711||October 31, 2002||Sartor et al.|
|20030097320||May 22, 2003||Gordon|
|20050182712||August 18, 2005||Angell|
|20060010093||January 12, 2006||Fan et al.|
|20060202012||September 14, 2006||Grano et al.|
|20070106580||May 10, 2007||Yang et al.|
|20070119919||May 31, 2007||Hogg et al.|
|20080091545||April 17, 2008||Jennings et al.|
|20080109392||May 8, 2008||Nandy|
|20090192957||July 30, 2009||Subramanian et al.|
|20090265211||October 22, 2009||May et al.|
|20100094765||April 15, 2010||Nandy|
Foreign Patent Documents
- Debar et al., “A Neural Network Component for an Intrusion Detection System”, Proceedings of the Computer Society Symposium on Research in Security and Privacy, vol. SYMP.13, May 4, 1992, 240-250.
- “U.S. Appl. No. 11/593,962 Notice of Allowance Mailed Sep. 15, 2009”, 10.
- “U.S. Appl. No. 11/593,962, Respoonse filed Jul. 24, 2009 to Non Final Office Action mailed Mar. 27, 2009”, 15 pgs.
- “U.S. Appl. No. 12/638,942, Non-Final Office Action mailed Apr. 9, 2010”, 2 pgs.
- “U.S. Appl. No. 12/638,942,—Notice of Allowance mailed Jul. 20, 2010”, 9.
- Devroye, L., “Any Discrimination Rule Can Have an Arbitrarily Bad Probability of Error for Finite Sample Size”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-4 , Issue: 2, (1982), 154-157.
- Kohler, M., et al., “On the Rate of Convergence of Local Averaging Plug-In Classification Rules Under a Margin Condition Under a Margin Condition, Kohler, M.; Krzyzak, A.;”, IEEE on Information Theory, 53(5), (2007), 1735-1742.
- Leung, A., et al., “On designing a flexible e-payment system with fraud detection capability”, IEEE International Conference on e-Commerce Technology, 2004. CEC 2004. Proceedings., (2004), 236-243.
- Martella, G., et al., “An Availability Model for Distributed Transaction Systems”, IEEE Transactions on Software Engineering, Volume: SE-11, Issue: 5, (1985), 483-491.
- Metwally, Ahmed, et al., “On Hit Inflation Techniques and Detection in Streams of Web Advertising Networks”, Distributed Computing Systems, 2007. ICDCS '07. 27th International Conference, (2007), 52-52.
- Montgomery, B. L, et al., “On decoding rules to minimize the probability of information bit errors”, IEEE Transactions on Information Theory, 34(4), (1988), 880-881.
- Rozsnyai, S, et al., “Solution architecture for detecting and preventing fraud in real time”, International Conference on Digital Information Management, 2007. ICDIM '07, vol. 1, (2007), 152-158.
- Venkataram, P., et al., “A Method of Fraud & Intrusion Detection for E-payment Systems in Mobile e-Commerce”, IEEE International Performance, Computing, and Communications Conference, 2007. IPCCC 2007., (2007), 395-401.
- “U.S. Appl. No. 11/593,962 Non-Final Office Action mailed Mar. 27, 2009”, OARN, 16 pgs.
- “U.S. Appl. No. 12/638,942, Response filed Jul. 8, 2010 to Non Final Office Action mailed Apr. 9, 2010”, 3 pgs.
Filed: Nov 4, 2010
Date of Patent: Nov 27, 2012
Patent Publication Number: 20110055078
Assignee: eBay, Inc. (San Jose, CA)
Inventor: Palash Nandy (San Francisco, CA)
Primary Examiner: Robert Niquette
Attorney: Schwegman Lundberg & Woessner, P.A.
Application Number: 12/939,936
International Classification: G06Q 40/00 (20120101);